01
Small, inspectable chunks
Each chunk covers one topic. Whole chunks preserve source meaning; budget packing skips an oversized chunk instead of cutting a sentence.
back to all labs
RAG Retrieval VisualizerLocal retrieval mode
Retrieval, with the decisions exposed.
Tune the retrieval policy. Watch the evidence change. Inspect every decision.
Retrieval policy
Changes apply instantly to the last search.
Uses a fixed public corpus on this site. No embeddings or external AI provider are called.
Showing evidence for “What AI engineering topics does Bondeth work on?”
- 01
Normalized terms
5 Normalized query terms
- 02
All sources ranked
4/7 sources matched
- 03
Cited context
3 sources kept
Context updated: 3 sources kept, 521 characters.
02 / Rank
6msSeven public chunks · one topic per chunk · English + Khmer
aiengineeringtopicsbondethwork
Inspect a source
[ai-focus] · In context
His AI interests include reliable structured outputs, retrieval augmented generation, RAG pipelines, agentic workflows, tool calling, and practical LLM evaluation.
aiengineering- Query terms matched
- 2/5
- Matching occurrences
- 3
- Source + citation size
- 195
80% term coverage + 5% per occurrence (up to 4), capped at 99%.
A keyword score, not a probability of correctness.
03 / Assemble
Context budget
521/700
Compared with baseline
33 sources kept
Same sources as baseline
Baseline: up to 3 matching sources, no score cutoff, 1,200 characters.
Evidence-backed preview
Exact source excerpts; no model-generated answer.
His AI interests include reliable structured outputs, retrieval augmented generation, RAG pipelines, agentic workflows, tool calling, and practical LLM evaluation.
Rithy Bondeth is a full stack developer and AI engineer based in Phnom Penh, Cambodia. He builds web, mobile, and intelligent software systems.
ហែម ឫទ្ធីបណ្ឌិត ជាវិស្វករសូហ្វវែរ និង AI នៅរាជធានីភ្នំពេញ។ គាត់អភិវឌ្ឍកម្មវិធីវែប កម្មវិធីទូរស័ព្ទ និងប្រព័ន្ធ AI។
Inspect the assembled context
Generation contract
Use only the cited context. Cite source IDs for factual claims. If the context does not support the question, say so.
[ai-focus] AI engineering focus His AI interests include reliable structured outputs, retrieval augmented generation, RAG pipelines, agentic workflows, tool calling, and practical LLM evaluation. [profile] Professional profile Rithy Bondeth is a full stack developer and AI engineer based in Phnom Penh, Cambodia. He builds web, mobile, and intelligent software systems. [khmer-profile] ប្រវត្តិរូបសង្ខេប ហែម ឫទ្ធីបណ្ឌិត ជាវិស្វករសូហ្វវែរ និង AI នៅរាជធានីភ្នំពេញ។ គាត់អភិវឌ្ឍកម្មវិធីវែប កម្មវិធីទូរស័ព្ទ និងប្រព័ន្ធ AI។
Why I built it this way
02
An honest retrieval baseline
Exact word matches make scoring inspectable and avoid matches such as “AI” inside “calling”. They miss synonyms and cross-language meaning. For production, I would evaluate hybrid retrieval against a labeled query set.
03
A deliberate evidence budget
A lower cutoff improves recall but admits weaker matches. A small budget keeps context focused but can omit useful evidence. The live trace exposes both tradeoffs.
These seven excerpts are curated for this demo. This is not semantic search, a measured accuracy benchmark, or a live LLM call.